Neural Networks

My main research interest are recurrent neural networks (RNN), which are particularly interesting for sequential and time-series data. Training RNNs with traditional methods can run into problems like slow convergence or oscillating weights. So called "Reservoir Computing" approaches avoid some of these by using larger, randomly fixed connectivity matrices as a dynamic, non-linear reservoir. This dynamic reservoir is driven by the input, and the task is to learn the best combination of internal dynamics to match the desired output. Some of our work investigates properties of reservoir computing approaches, and finding the right measures to do so (see below). We also look at ways to improve training methods, both for reservoir computing approaches or the traditional methods.

RNN Applications

[Nano Neurons Icon] Nano Neurons: This is a current project at CSIRO in the intersection between material science and computer science, where we investigate use of small particles to create materials that can react to changes in the environment, or even process information. Some of our work has been published in the journal Nano Communication Networks.

[robot arm]Inverse Dynamics Learning for a Robot Arm with Artificial Muscles: Robots with artificial muscles and many degrees of freedom are versatile and compliant - this makes them ideal for use in environments close to humans. Control of these robots, however, can be challenging even for machine learning methods because of the many non-linearities, delays in actuation, and unknown parameters such as friction. We propose a novel method that uses RNN, Gaussian Processes, and Goal Babbling to learn control of redundant robot arms with artificial muscles online from scratch using only the position of the end effector, without using any joint positions, accelerations or an analytical model of the system or the environment. There is a project overview web page and electronic poster, and also is also a paper that has been published at Robotics: Systems and Science (R:SS 2012).

CSIRO Sensor Nets Intelligent Information Processing in Sensor Networks Sensor networks are distributed systems used for environmental monitoring, for example to detect interesting or abnormal conditions. We investigated approaches to detect anomalies in concentrations of gases in underground coal mines [WLO+08], [OWP08], and methods to train recurrent neural networks distributed over an entire sensor network [Obst09d], [Obst13b].

causalityTime Series Causality Inference using Echo State Networks
One potential strength of recurrent neural networks (RNNs) is their theoretical ability to find a connection between cause and consequence in time series in an constraint-free manner, that is without the use of explicit probability theory. Here, we use echo state networks for this purpose. Our approach learns probabilities explicitly using an online learning procedure (recursive least squares) and echo state networks.

Training Recurrent Neural Networks

[Neurons]Taming the Reservoir: Feedforward Training for Recurrent Neural Networks
As mentioned above, traditional training methods for RNN have their problems. Performance of reservoir computing methods, on the other hand, is quite variable due to the randomly connected hidden layer. We came up with an approach that makes use of efficient feedforward training methods to initialise the recurrent reservoir, and performs better than echo state networks for some time series prediction tasks. Though our work considers reservoir computing approaches, the same approach could be used to initialise weights for traditional methods like backpropagation through time (BPTT) [OR12].

information transferImproving Recurrent Neural Network Performance Using Transfer Entropy
For randomly initialised reservoirs, we are interested in methods that improve the reservoir locally, self-organised way: this means that this part of the training can be distributed very easily. Many plasticity mechanisms work like this, but they are rarely influenced by the overall learning goal. Using information transfer, we adapt the memory in the system dependent on the desired output. More...

Properties of Recurrent Reservoirs & Measures

HFSP coverInitialization and Self-Organized Optimization of Recurrent Neural Network Connectivity
We study a general network initialization method using permutation matrices and derive a new unsupervised learning rule based on intrinsic plasticity (IP). The IP based learning uses only local learning, and its aim is to improve network performance in a self-organized way. We show that networks with permutation matrices for the reservoir connectivity have much more persistent memory than the other methods, but are also able to perform highly non-linear mappings. We also show that IP based on sigmoid transfer functions is limited concerning the output distributions that can be achieved. See our paper [BOMA09b] for more details.

Edge of ChaosInformation Processing in Echo State Networks at the Edge of Chaos
We investigate information processing in randomly connected recurrent neural networks. It has been shown previously that the computational capabilities of these networks are maximized when the recurrent layer is close to the border between a stable and an unstable dynamics regime, the so called edge of chaos. The reasons, however, for this maximized performance are not completely understood. We adopt an information-theoretical framework and are for the first time able to quantify the computational capabilities between elements of these networks directly as they undergo the phase transition to chaos. Specifically, we present evidence that both information transfer and storage in the recurrent layer peak close to this phase transition, providing an explanation for why guiding the recurrent layer towards the edge of chaos is computationally useful. For more details, see [BOL+12].

[Neurons]Guided Self-Organization of Input-Driven Recurrent Neural Networks
Neuroscientists increasingly look for theoretical frameworks that could help explain the data recorded from the brain, and to make the enormous task more manageable. This is evident, for instance, through the funding of the billion-dollar "Human Brain Project", amongst others. Mathematical techniques from graph and information theory, control theory, dynamical and complex systems (Sporns 2011), statistical mechanics (Rolls and Deco 2010), as well as machine learning and computer vision (Seung 2012; Hawkins and Blakeslee 2004), have provided new insights into brain structure and possible function, and continue to generate new hypotheses for future research. We review some of these techniques applicable to Recurrent Neural Network research in [OB14]. For our work in particular, we are interested in quantifying information storage, information modification, and information transfer as characteristic elements of computation. Although these quantities are defined for autonomous dynamical systems, they do not distinguish between the system itself, and the effects the input has on the system. We propose an input-corrected measure of information storage, important for heavily input-driven systems like artificial neural networks to abstract from specific benchmarks, or for brain networks, where intervention is difficult. A preprint is available here [OBSA13b].